基于多层全连接神经网络的湖南省流域面雨量预报  

Areal Rainfall Forecast over Hunan Drainage Basin Using Multi-layer Fully Connected Neural Network

作　　者：谭诗琪 范嘉智 廖春花 罗潇 龙晓琴 罗立军 卞一飞 Tan Shiqi;Fan Jiazhi;Liao Chunhua;Luo Xiao;Long Xiaoqin;Luo Lijun;Bian Yifei(Hunan Meteorological Service Center,Changsha 410118,China;Hunan Key Laboratory of Meteorological Disaster Prevention and Mitigation,Changsha 410118,China;China Meteorological Administration Training Centre Hunan Branch,Changsha 410125,China;Hunan Wuling Power Technology Corporation Ltd.,Changsha 410029,China;Service Center of Hunan Provincial Meteorological Bureau,Changsha 410118,China)

机构地区：[1]湖南省气象服务中心,长沙410118 [2]气象防灾减灾湖南省重点实验室,长沙410118 [3]中国气象局气象干部培训学院湖南分院,长沙410125 [4]湖南五凌电力科技有限公司,长沙410029 [5]湖南省机关服务中心,长沙410118

出　　处：《气象与环境科学》2024年第5期62-69,共8页Meteorological and Environmental Sciences

基　　金：湖南省气象局第三期业务能力建设项目(NLJS09);2024年度湖南省自然科学基金项目(2024JJ9167)。

摘　　要：研究基于ECMWF、JMA东亚地区再分析资料、OCF、湖南省智能网格预报0.5°格点产品、华南区域数值天气预报模式及华东区域数值天气预报模式产品,利用多层全连接神经网络(MFCNN)构建模型,预测未来24 h、3 h、1 h面雨量数据,采用平均绝对误差(MAE)、均方根误差(RMSE)和决定系数(R^(2)),对2020年湖南大型水库流域面雨量MFCNN模型预报效果进行检验评估及与各家模式预报效果对比分析。结果表明:MFCNN模型对24 h、3 h、1 h面雨量的预报效果(MAE、RMSE、R^(2))均优于各模式的效果,且随时间分辨率的提升,模型预测效果相对于各模式提升明显。误差系数表明,MFCNN模型预报偏差在湘江流域及洞庭湖的最小,在沅水中游、资水上游的居中,在澧水流域、资水下游及沅水上、下游的最大。该模型捕捉面雨量动态变化的能力在洞庭湖、澧水上游、沅水中游最强,在湘江流域的次之,在沅水上、下游及资水下游、澧水下游的最弱。Based on ECMWF,JMA reanalysis data in East Asia,OCF,0.5°grid products of intelligent grid forecast in Hunan Province,numerical weather forecast model products in South China and East China,this research constructs a model by using multilayer fully connected neural network(MFCNN)to predict the future 24 h,3 h and 1 h areal rainfall.The average absolute error(MAE),root mean square error(RMSE)and determination coefficient(R^(2))are used to test and evaluate the prediction effect of MFCNN model for regional rainfall of Hunan major reservoir drainage basin in 2020,and the forecast effect is compared with the effects of other models.The results show that the MFCNN model has better forecasting effects(MAE,RMSE,R^(2))of 24 h,3 h and 1 h areal precipitation than other models.With the improvement of time resolution,the prediction effect of MFCNN model becomes even better.The error coefficient shows that the precipitation predicted by the MFCNN model has the smallest deviation in Xiangjiang River and Dongting Lake,the moderate deviation in middle reach of Yuanshui River and the upper reach of Zishui River,and the largest deviation in Lishui River basin,the lower reach of Zishui River and the upper and lower reaches of Yuanshui River.It has been found that the MFCNN model has the strongest ability to capture the dynamic variation of areal rainfall in Dongting Lake,upper reach of Lishui River and middle reach of Yuanshui River,the second in Xiangjiang River basin,and the weakest in upper and lower reaches of Yuanshui River,the low reach of Zishui River and the lower reach of Lishui River.

关 键 词：面雨量预报 数值模式 多层全连接神经网络模型 

分 类 号：P339[天文地球—水文科学]